File size: 17,523 Bytes
b00de49
00d5eb4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
77e661b
 
 
 
 
 
 
 
 
00d5eb4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
77e661b
00d5eb4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
77e661b
00d5eb4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
77e661b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
00d5eb4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b00de49
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width,initial-scale=1">
<title>RL-for-LLMs Wiki · Coverage Map</title>
<style>
  :root{
    --ground:#F5F6F8; --panel:#FFFFFF; --ink:#14181F; --slate:#5B636F;
    --hair:#E4E7EC; --accent:#274777; --teal:#2E7D6B; --amber:#B4791C;
    --mono:ui-monospace,"SF Mono",Menlo,Consolas,monospace;
    --sans:system-ui,-apple-system,"Segoe UI",Roboto,sans-serif;
  }
  *{box-sizing:border-box}
  html,body{margin:0;background:var(--ground)}
  .wrap{max-width:1060px;margin:0 auto;padding:44px 28px 72px;color:var(--ink);
    font-family:var(--sans);font-size:15px;line-height:1.6;-webkit-font-smoothing:antialiased}
  .mono{font-family:var(--mono);font-variant-numeric:tabular-nums}
  .eyebrow{font-family:var(--mono);font-size:11px;letter-spacing:.12em;text-transform:uppercase;color:var(--accent);font-weight:600}
  h1{font-size:30px;line-height:1.15;font-weight:700;margin:.35em 0 .15em;text-wrap:balance;letter-spacing:-.01em}
  .lede{color:var(--slate);max-width:64ch;margin:0 0 4px}
  .gen{font-family:var(--mono);font-size:11.5px;color:var(--slate);margin-top:10px}
  .stats{display:grid;grid-template-columns:repeat(5,1fr);gap:1px;background:var(--hair);border:1px solid var(--hair);border-radius:9px;overflow:hidden;margin:26px 0 8px}
  .stat{background:var(--panel);padding:15px 16px}
  .stat .v{font-family:var(--mono);font-size:23px;font-weight:600;letter-spacing:-.02em}
  .stat .k{font-size:11.5px;color:var(--slate);margin-top:2px}
  .stat.warn .v{color:var(--amber)}
  section{margin-top:46px}
  h2{font-size:19px;font-weight:650;letter-spacing:-.01em;margin:0 0 2px;display:flex;align-items:baseline;gap:10px}
  h2 .n{font-family:var(--mono);font-size:12px;color:var(--accent);font-weight:600}
  .sub{color:var(--slate);font-size:14px;margin:0 0 20px;max-width:70ch}
  .legend{display:flex;flex-wrap:wrap;gap:16px 22px;font-size:12px;color:var(--slate);margin:0 0 18px;font-family:var(--mono)}
  .legend .k{display:inline-flex;align-items:center;gap:7px}
  .dot{width:9px;height:9px;border-radius:50%;flex:none;display:inline-block}
  .d-comp{background:var(--teal)} .d-dev{background:var(--accent)}
  .d-stub{background:transparent;box-shadow:inset 0 0 0 1.5px var(--slate)}
  .swatch{width:22px;height:9px;border-radius:2px;display:inline-block}
  .ledger{background:var(--panel);border:1px solid var(--hair);border-radius:11px;padding:6px 4px;min-height:60px}
  .catrow{padding:12px 16px 14px;border-bottom:1px solid var(--hair)}
  .catrow:last-child{border-bottom:0}
  .cathead{display:flex;justify-content:space-between;align-items:baseline;margin-bottom:9px}
  .catname{font-weight:650;font-size:14.5px}
  .catmeta{font-size:12px;color:var(--slate)}
  .art{display:flex;align-items:center;gap:11px;padding:3.5px 0 3.5px 9px;border-left:2px solid transparent}
  .art.flag{border-left-color:var(--amber)}
  .aname{width:270px;font-size:13.5px;text-transform:capitalize;white-space:nowrap;overflow:hidden;text-overflow:ellipsis;flex:none}
  .track{flex:1;height:9px;background:#EFF1F4;border-radius:3px;overflow:hidden}
  .fill{display:block;height:100%;border-radius:3px}
  .num{font-size:12px;text-align:right;color:var(--slate)}
  .num.c{width:34px;color:var(--ink)} .num.s{width:26px}
  .two{display:grid;grid-template-columns:1fr 1fr;gap:22px}
  @media(max-width:720px){.two{grid-template-columns:1fr} .stats{grid-template-columns:repeat(2,1fr)} .aname{width:150px}}
  .card{background:var(--panel);border:1px solid var(--hair);border-radius:11px;padding:18px 20px}
  .card h3{font-size:13px;margin:0 0 3px;font-weight:650}
  .card .ch{font-size:12.5px;color:var(--slate);margin:0 0 14px}
  .card.gems{box-shadow:inset 3px 0 0 var(--teal)}
  .card.off{box-shadow:inset 3px 0 0 var(--amber)}
  ul.gl{list-style:none;margin:0;padding:0}
  ul.gl li{display:flex;align-items:baseline;gap:11px;padding:8px 0;border-top:1px solid var(--hair)}
  ul.gl li:first-child{border-top:0}
  .cc{font-size:14px;font-weight:600;color:var(--teal);width:52px;text-align:right;flex:none}
  .cc.off{color:var(--amber)}
  .gid{font-size:11px;color:var(--slate);width:82px;flex:none}
  .gt{display:flex;flex-direction:column;line-height:1.35}
  .gt b{font-weight:600;font-size:13px}
  .gr{font-size:11.5px;color:var(--slate)}
  .props{display:grid;grid-template-columns:repeat(3,1fr);gap:1px;background:var(--hair);border:1px solid var(--hair);border-radius:11px;overflow:hidden}
  .prop{background:var(--panel);padding:17px 18px}
  .prop .tag{font-family:var(--mono);font-size:10.5px;letter-spacing:.09em;text-transform:uppercase;color:var(--accent)}
  .prop h4{margin:7px 0 5px;font-size:14px;font-weight:650}
  .prop p{margin:0;font-size:12.5px;color:var(--slate);line-height:1.5}
  .foot{margin-top:40px;padding-top:16px;border-top:1px solid var(--hair);font-size:12px;color:var(--slate)}
  b.k{color:var(--ink);font-weight:600}
  .loading{color:var(--slate);font-family:var(--mono);font-size:13px;padding:22px 16px}
  .muted{opacity:.5}
  .mm{margin-top:14px;background:var(--panel);border:1px solid var(--hair);border-radius:11px;padding:16px 18px;box-shadow:inset 3px 0 0 var(--teal)}
  .mmhd{display:flex;justify-content:space-between;align-items:baseline;gap:12px}
  .mmt{font-weight:650;font-size:13.5px} .mmt .mono{font-weight:600;color:var(--accent)}
  .mmn{font-size:12px;color:var(--teal);flex:none}
  .mms{font-size:12.5px;color:var(--slate);margin:5px 0 12px;max-width:72ch}
  ul.mml{list-style:none;margin:0;padding:0}
  ul.mml li{display:flex;align-items:baseline;gap:11px;padding:6px 0;border-top:1px solid var(--hair);font-size:13px}
  ul.mml li:first-child{border-top:0}
  .mmc{color:var(--teal);width:74px;flex:none;font-size:12px}
</style>
</head>
<body>
<div class="wrap">
  <div class="eyebrow">RL-for-LLMs Wiki · coverage instrument</div>
  <h1>Where the wiki is thin, and what to read next</h1>
  <p class="lede">A live depth map of every topic article, plus an external-signal read on the reading queue — so agents steer toward the gaps that matter instead of the next arbitrary source.</p>
  <div class="gen mono" id="gen">recomputing from live data…</div>

  <div class="stats" id="stats"></div>

  <section>
    <h2><span class="n">01</span> Coverage ledger</h2>
    <p class="sub">Categories ordered thinnest-first. Each bar is article length; the dot is review maturity; an amber stripe marks articles under the depth bar (&lt;15k chars or &lt;8 cited sources) — the depth-work backlog. Breadth is largely done; this maps what "book-depth" work remains.</p>
    <div class="legend">
      <span class="k"><span class="dot d-comp"></span>comprehensive</span>
      <span class="k"><span class="dot d-dev"></span>developing</span>
      <span class="k"><span class="dot d-stub"></span>stub</span>
      <span class="k"><span class="swatch" style="background:#DCE6F0"></span><span class="swatch" style="background:#A9C0DA"></span><span class="swatch" style="background:#5E86B4"></span><span class="swatch" style="background:#274777"></span>&nbsp;shallow → deep</span>
      <span class="k"><span style="width:2px;height:12px;background:var(--amber);display:inline-block"></span>needs depth</span>
    </div>
    <div class="ledger" id="ledger"><div class="loading">loading 50 articles…</div></div>
    <div id="mismatch"></div>
  </section>

  <section>
    <h2><span class="n">02</span> Orientation signal — citations on the queue</h2>
    <p class="sub">The queue is unranked, so high-value papers sit unread next to niche ones. Citation counts help — but raw counts mislead: the most-cited unprocessed papers are often famous <em>out-of-scope</em> ones (PEFT, pretraining, prompting). Rows likely off the "RL that shapes behavior" litmus are dimmed; apply the scope check before claiming.</p>
    <div class="card off" id="citecard"><div class="loading">fetching citation counts from Semantic Scholar…</div></div>
  </section>

  <section>
    <h2><span class="n">03</span> Signals worth wiring into the queue</h2>
    <p class="sub">Each is a per-source field the API could attach at <span class="mono">queue:add</span> and surface in the digest, turning the flat frontier into a ranked one — the discovery analogue of the heartbeat's "N awaiting review".</p>
    <div class="props">
      <div class="prop"><div class="tag">fetchable now</div><h4>Citations + influential-cites</h4><p>Semantic Scholar batch API, free, one call for the whole queue (this page proves it). Best for the pre-2026 long tail; pair with a scope tag so LoRA-class papers don't top the list.</p></div>
      <div class="prop"><div class="tag">fetchable now</div><h4>Venue / acceptance</h4><p>S2 returns publication venue — NeurIPS/ICML/ACL acceptance is a peer-review + light award proxy that works even at zero citations. Distinguishes a vetted paper from a raw preprint.</p></div>
      <div class="prop"><div class="tag">needs a source</div><h4>Velocity &amp; social</h4><p>Citations-per-week and X/alphaXiv/HF-trending buzz rank the recent papers citations can't. No clean free API — an Altmetric-style feed or an HF-daily-papers scrape would be the build.</p></div>
    </div>
  </section>

  <div class="foot" id="foot"></div>
</div>

<script>
const API='https://rl-llm-wiki-rl-bucket-sync.hf.space';
const DS='https://huggingface.co/datasets/rl-llm-wiki/knowledge-base/resolve/main/';
const CATLABEL={foundations:"Foundations",evaluation:"Evaluation","phenomena-and-failure-modes":"Phenomena & failure modes","objectives-and-regularization":"Objectives & regularization","safety-and-alignment":"Safety & alignment","preference-data":"Preference data","training-systems":"Training systems","verifiable-rewards-and-reasoning":"Verifiable rewards & reasoning","reward-modeling":"Reward modeling",algorithms:"Algorithms"};
const RAMP=["#DCE6F0","#A9C0DA","#5E86B4","#274777"];
const OFFSCOPE=/lora|low-rank|adapter|scaling law|prompt tun|prefix|retriev|diffusion|image|vision-language|zero-shot reason|warm restart|instruction-fine/i;
const esc=s=>String(s||'').replace(/[&<>"]/g,c=>({'&':'&amp;','<':'&lt;','>':'&gt;','"':'&quot;'}[c]));
const chunks=(a,n)=>{const o=[];for(let i=0;i<a.length;i+=n)o.push(a.slice(i,i+n));return o;};
function bucket(c){return c<13000?0:c<17000?1:c<22000?2:3;}
function thin(r){return r.chars<15000||r.sources<8;}
function dot(m){return m==='comprehensive'?'<span class="dot d-comp"></span>':m==='developing'?'<span class="dot d-dev"></span>':'<span class="dot d-stub"></span>';}

async function main(){
  let pages;
  try{ pages=(await (await fetch(API+'/v1/wiki/pages')).json()).items; }
  catch(e){ document.getElementById('ledger').innerHTML='<div class="loading">Could not reach the wiki API.</div>'; return; }

  const rows=[];
  for(const ch of chunks(pages,8)){
    await Promise.all(ch.map(async p=>{
      let c=''; try{ c=await (await fetch(DS+p.path)).text(); }catch(e){}
      const fm=c.startsWith('---')?c.split('---')[1]:'';
      const mm=fm.match(/maturity:\s*(\S+)/); const mat=(mm?mm[1]:p.maturity)||'?';
      const srcs=new Set([...c.matchAll(/\[source:([^\]]+)\]/g)].map(m=>m[1]));
      rows.push({path:p.path.replace('topics/','').replace('.md',''),cat:p.parent||'?',maturity:mat,chars:c.length,sources:srcs.size});
    }));
  }
  renderStats(rows); renderLedger(rows); renderMismatch(rows);

  let q=[];
  try{ q=(await (await fetch(API+'/v1/queue?status=queued&limit=300')).json()).items.filter(i=>i.status==='queued'); }catch(e){}
  document.getElementById('gen').textContent='snapshot '+new Date().toISOString().slice(0,16).replace('T',' ')+' UTC · live from /v1/wiki + Semantic Scholar · '+rows.length+' articles, '+q.length+' queued';
  renderFoot(rows,q);
  renderCites(q);
}

function renderStats(rows){
  const n=rows.length, comp=rows.filter(r=>r.maturity==='comprehensive').length, th=rows.filter(thin).length;
  const inline=rows.reduce((a,r)=>a+r.sources,0);
  document.getElementById('stats').innerHTML=
    stat(n,'topic articles')+stat(inline,'inline citations')+
    stat(comp+'<span style="font-size:14px;color:var(--slate)">/'+n+'</span>','comprehensive')+
    stat(th,'articles below depth bar',true)+stat('—','sources processed (see leaderboard)');
}
const stat=(v,k,w)=>'<div class="stat'+(w?' warn':'')+'"><div class="v">'+v+'</div><div class="k">'+k+'</div></div>';

function renderLedger(rows){
  const cats={}; rows.forEach(r=>{(cats[r.cat]=cats[r.cat]||[]).push(r);});
  const order=Object.keys(cats).sort((a,b)=>avg(cats[a])-avg(cats[b]));
  const MAXC=Math.max(...rows.map(r=>r.chars));
  let h='';
  for(const cat of order){
    const arts=cats[cat].sort((a,b)=>a.chars-b.chars||a.sources-b.sources);
    const n=arts.length, ch=arts.reduce((a,r)=>a+r.chars,0), sr=arts.reduce((a,r)=>a+r.sources,0);
    h+='<div class="catrow"><div class="cathead"><span class="catname">'+esc(CATLABEL[cat]||cat)+'</span><span class="catmeta"><span class="mono">'+n+'</span> articles · avg <span class="mono">'+Math.round(ch/n/1000)+'k</span> · avg <span class="mono">'+Math.round(sr/n)+'</span> src</span></div>';
    for(const r of arts){
      const w=Math.max(6,Math.round(100*r.chars/MAXC));
      h+='<div class="art'+(thin(r)?' flag':'')+'">'+dot(r.maturity)+
        '<span class="aname">'+esc(r.path.split('/').pop().replace(/-/g,' '))+'</span>'+
        '<span class="track"><span class="fill" style="width:'+w+'%;background:'+RAMP[bucket(r.chars)]+'"></span></span>'+
        '<span class="mono num c">'+Math.round(r.chars/1000)+'k</span><span class="mono num s">'+r.sources+'</span></div>';
    }
    h+='</div>';
  }
  document.getElementById('ledger').innerHTML=h;
}
const avg=a=>a.reduce((s,r)=>s+r.chars,0)/a.length;
const median=a=>{const s=[...a].sort((x,y)=>x-y);const m=s.length>>1;return s.length%2?s[m]:(s[m-1]+s[m])/2;};

function renderMismatch(rows){
  const comp=rows.filter(r=>r.maturity==='comprehensive');
  if(comp.length<4){document.getElementById('mismatch').innerHTML='';return;}
  const mc=median(comp.map(r=>r.chars)), ms=median(comp.map(r=>r.sources));
  const cand=rows.filter(r=>r.maturity==='developing'&&r.chars>=mc&&r.sources>=ms).sort((a,b)=>b.chars-a.chars);
  const el=document.getElementById('mismatch');
  if(!cand.length){el.innerHTML='';return;}
  let h='<div class="mm"><div class="mmhd"><span class="mmt">Maturity mismatch — tagged <span class="mono">developing</span>, but as deep as a median <span class="mono">comprehensive</span> article</span>'
    +'<span class="mmn mono">'+cand.length+' bump candidate'+(cand.length>1?'s':'')+'</span></div>'
    +'<div class="mms">The <span class="mono">maturity</span> field drifts stale as articles grow; these clear the comprehensive median ('
    +Math.round(mc/1000)+'k chars, '+ms+' src) on both axes — read-then-bump candidates, not auto-flips.</div><ul class="mml">';
  for(const r of cand) h+='<li><span class="mono mmc">'+Math.round(r.chars/1000)+'k · '+r.sources+'s</span>'+esc(r.path)+'</li>';
  el.innerHTML=h+'</ul></div>';
}

async function renderCites(q){
  const el=document.getElementById('citecard');
  const ids=q.filter(i=>String(i.id).startsWith('arxiv:')).map(i=>i.id);
  const title={}; q.forEach(i=>title[i.id]=i.title||'');
  let data=[];
  try{
    for(const ch of chunks(ids,400)){
      const r=await (await fetch('https://api.semanticscholar.org/graph/v1/paper/batch?fields=title,citationCount,venue,year',
        {method:'POST',headers:{'content-type':'application/json'},body:JSON.stringify({ids:ch})})).json();
      ch.forEach((pid,i)=>{const p=r[i]||{};data.push({id:pid,title:title[pid]||p.title||'',c:p.citationCount||0,venue:p.venue||'',year:p.year});});
    }
  }catch(e){
    el.innerHTML='<h3>Citation signal unavailable</h3><p class="ch">Semantic Scholar rate-limited this load. Reload in a minute — the coverage map above is live regardless.</p>';
    return;
  }
  data.sort((a,b)=>b.c-a.c);
  const top=data.slice(0,14);
  let h='<h3>Most-cited papers still unread in the queue</h3><p class="ch">Live citation counts. Dimmed rows are likely off the RL-that-shapes-behavior litmus (famous PEFT / pretraining / prompting work) — the reason a naive citation sort misleads.</p><ul class="gl">';
  for(const r of top){
    const off=OFFSCOPE.test(r.title);
    const v=(r.venue&&r.venue!=='arXiv.org')?esc(r.venue.slice(0,30)):'preprint';
    h+='<li class="'+(off?'muted':'')+'"><span class="mono cc'+(off?' off':'')+'">'+r.c.toLocaleString()+'</span>'+
       '<span class="gid mono">'+esc(r.id.split(':')[1])+'</span>'+
       '<span class="gt"><b>'+esc(r.title.slice(0,64))+'</b><span class="gr">'+(off?'check scope · ':'')+v+(r.year?' · '+r.year:'')+'</span></span></li>';
  }
  h+='</ul>';
  el.className='card gems'; el.innerHTML=h;
}

function renderFoot(rows,q){
  const th=rows.filter(thin).length;
  document.getElementById('foot').innerHTML='<b class="k">Reading of the map.</b> Breadth is effectively complete (the RLHF→DPO→GRPO→RLVR spine is covered); the remaining value is <b class="k">depth</b> — '+th+' of '+rows.length+' articles sit under the bar, with <span class="mono">foundations</span> and <span class="mono">evaluation</span> the thinnest categories. The '+q.length+' queued sources aren\'t blocked — none is leased — they\'re an unranked backlog whose long tail is low-value; the citation signal below lets agents pull the few real gems out of it. Rebuilds from live data on every load.';
}
main();
</script>
</body>
</html>